Multilingual IR with Machine-Translated FAQ

[stedomedo]

Building multilingual models (zero-shot, transfer learning, etc.) takes time.

So, in the meantime, as stated in #2 , we could machine-translate FAQs from English into other languages and add them to the search cluster, so that they can be retrieved at foreign language input. Translations in the background don't need to be perfect, but sufficient for retrieval (adequacy before fluency/grammar).

TODOs:

• Scrape the English FAQ from data/scrapers repo
• Build machine-translator tool (e.g. with https://pypi.org/project/googletrans/)
• Translate some samples to check quality
• Translate all English FAQ
• Add data to ESC with columns: language, original_english_doc, is_machine_translated

[tholor]

Great idea. @stedomedo! Did I get this right, that we would still need language-specific models for question similarity with this approach?

Would it be an alternative to translate the user question live to English and then to the matching with our FAQs? With that approach we could easily leverage English models for question similarity.

[stedomedo]

Yes, that's an option.
Query translation quality could suffer though from short lengths.
I'm currently exploring translation quality.
Thanks!

[stedomedo]

The googletrans lib does not work reliably, so I made a free trial account on MS Azure, also because they offer up to 2M characters of translation for free per month.

Here is the English FAQ data including columns for Arabic:
https://github.com/stedomedo/COVID-QA/blob/auto_translators/data/faqs/MT_ar_faq_covidbert.csv

[stedomedo]

And the MS translator:
https://github.com/stedomedo/COVID-QA/blob/auto_translators/data/translators/ms_translate.py

MS Translator is supposed to be quite good for Arabic. For other languages, Google or DeepL are better options (afaik they don't offer free credits)

Checking which real-time translation option is best to use, incl. budget-wise

[stedomedo]

@tholor @Timoeller I have a question on the (desired) search workflow.
Is it:
user query -> match query to question with BERT -> search with elastic (tfidf, bm25)
?

So could a multilingual workflow be like this:
query -> detect lang
-> if EN -> match query to question with BERT -> search with elastic (tfidf, bm25)
-> if AR -> search directly with elastic (tfidf, bm25)
?
In this case, no multilingual-BERT or other-language-BERT or real-time translation would be needed.

[Timoeller]

Good points.

Can you create a PR with the translation and the script for doing so? I would merge it to have this functionality in the repo.

About the language detection and the switch between bert + ES and only ES: we could implement it this way if multilingual isnt working well for other languages.

Do you have experience with language detection and could write a script for this, so we can integrate this into the backend? We need lang detection there anyways, because we want to adjust output texts like "source" "cateogry" etc. The script should be rather efficient, since this will limit response time...

[stedomedo]

One idea for "simple" transfer learning:
In Machine Translation this technique is commonly used when you have a low resource language. Basically, you build a model for language Y on top of the model for language X by just continuing the training (1-2 epochs) with the Y language data. Vocabs would need to be pooled on all languages though.
This could work for small data sizes or/and maybe also machine-translated texts.

[Timoeller]

That is exactly the idea! : )
With multilingual models like mBert or XLM-R this "zero shot learning" is easily possible because the vocab is already in one pool for all supported languages.
See e.g. Table 1 or 3 in XLM-R paper for zero shot transfer.

So if we train a multilingual model in Sentence Bert on Quora, we will also be able to match all other languages - hopefully with good performance 💃

[ViktorAlm]

You are probably aware of these datasets but heres some multilingual similarity data. I have a NMT model for english->swedish if you want me to I could NMT and add some data for better performance on scandinavian languages.

https://github.com/google-research-datasets/paws
https://www.nyu.edu/projects/bowman/xnli/